Control Oriented Reinforcement Learning: A Survey of Recent Progress and Applications

ABSTRACT Modern control systems are expected not only to pursue optimality, but also to dynamically adapt to varying environments. Bridging the gap between adaptive control, optimal control, and data‐driven learning control, reinforcement learning has emerged as a computationally efficient approach to achieve adaptive optimal control of systems. This paper surveys both theoretical advancements and practical applications of control oriented reinforcement learning methods, especially adaptive dynamic programming (ADP). We discuss recent progress of ADP in several key control disciplines, including optimal control, robust control, event‐triggered control, distributed control and safe control; as well as real‐world applications of ADP in various scenarios such as unmanned vehicles, power systems, intelligent transportation, robot manipulators, and motors.